Today's edition highlights systems that build their own infrastructure, from agents mapping their own skill graphs to teachers writing reusable instructions for real-time execution.
The loop: A coding agent identifies persistent failures in its executor model and generates its own fine-tuning data to improve that executor for future tasks.
An on-demand recursive self-improvement framework allows an embodied agent to attribute failures and fine-tune its own executor using self-collected demonstrations.
The loop: An LLM agent proposes and commits execution-backed edges to its own skill graph as it works, improving its ability to retrieve and sequence skills for future tasks.
A retrieval framework exposes a typed skill graph to an LLM agent as a callable interface, allowing the agent to dynamically update the graph with new execution-backed relationships.
The loop: A slow teacher model studies failures in paused environments and writes reusable instructions that improve a fast student model's performance in real-time tasks.
A real-time benchmark evaluates agents while games continue running during inference, paired with an offline self-improvement method where a teacher model writes instructions to speed up a student model.
The loop: A learned proposer explores an environment to generate tasks and verifiable rewards, which are used in alternating training to improve both the solver and the proposer itself.
A self-play framework jointly generates tasks and fine-grained verifiable rewards through environment interaction to train agents on long-horizon tasks without human annotations.
The loop: A terminal agent recursively rewrites its own verified trajectories and uses them for policy refinement, iteratively updating its weights across multiple stages.
A recursive self-improvement method alternates self-rewriting of verified trajectories with policy refinement to improve terminal agents on long-horizon tasks.
The loop: The Zephon data loader improves the training process of models in TorchTitan, which makes the resulting models better or more efficiently trained.
Why Zephon? The data that a model sees during training is one of the most powerful levers we have for influencing how well it performs. Much of the day-to-day work... The post Bringing Elastically Deterministic Data Loading to TorchTitan...
